CED-Net: contextual encoder-decoder network for 3D face reconstruction

This paper proposes a contextual encoder-decoder network to regress UV position map for 3D face reconstruction, named CED-Net. Specifically, the CED-Net introduces contextual information both in the shape and feature level, which is served as the crucial information. Firstly, shape context relationship is considered as the loss function constraint, in which the intrinsic Euclidean norm and vector angle similarity are computed for each contextual vector. Furthermore, to add the contextual information at the feature level, the local feature correlation modulator is incorporated into the network's center section, which directs the network to capture the relationship between facial features from a spatial perspective. The heatmaps of the learned features demonstrate that the proposed CED-Net can pay attention to the comprehensive facial regions for reconstruction. The quantitative and qualitative experiments conducted on AFLW2000-3D and AFLW-LPFA show that the proposed method achieves superior results in both benchmarks.